Enhance Regional Wall Segmentation by Style Transfer for Regional Wall Motion Assessment
Kaikai Liu (Northwest A&F University), Yiyu Shi (University of Notre Dame), Jian Zhuang (Guangdong Provincial People's Hospital), Meiping Huang (Guangdong Provincial People's Hospital), Hongwen Fei (Guangdong Provincial People's Hospital), Boyang Li (Meta), Jin Hong (Guangdong Provincial People's Hospital), Qing Lu (University of Notre Dame), Erlei Zhang (Northwest A&F University), Xiaowei Xu (Guangdong Provincial People's Hospital)*
The 34th British Machine Vision Conference
Abstract
Regional wall motion assessment is critical in the diagnosis of coronary artery diseases and is usually performed using echocardiography in clinical practice. As manual assessment of regional wall motion is time-consuming and requires strong expertise, various automated methods have been proposed in which regional wall segmentation is the key step for detailed diagnosis and treatment. Generally, the existing methods suffer from low segmentation performance. In addition, there are no publicly available datasets which also introduces difficulties to related research in the community. In this paper, we propose to enhance regional wall segmentation by style transfer for regional wall motion assessment. The motivation is that echocardiography images in 2D mode are inexpensive but the contrast is low while echocardiography images in myocardial contrast echocardiography (MCE) mode and left ventricle opacification (LVO) mode are invasive and expensive but with high contrast, so style transfer may help enhance the contrast of echocardiography images in 2D mode. Specifically, we first transfer echocardiography images in 2D mode to echocardiography images in LVO or MCE mode using style transfer. Then, nnU-Net is adopted for regional wall segmentation. Note that the two networks are trained in an end-to-end manner. For evaluation, we collected a dataset of 198 patients each with three views (including A2C, A3C, and A4C) in three modes (including 2D mode, MCE mode, and LVO mode). Experimental results show that our framework can improve the Dice score by 7.53% compared with existing works. However, the Dice score is still below 60%, leaving much room for further improvement. The dataset and code in our study are released to the public.Citation
@inproceedings{Liu_2023_BMVC,
author = {Kaikai Liu and Yiyu Shi and Jian Zhuang and Meiping Huang and Hongwen Fei and Boyang Li and Jin Hong and Qing Lu and Erlei Zhang and Xiaowei Xu},
title = {Enhance Regional Wall Segmentation by Style Transfer for Regional Wall Motion Assessment},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year = {2023},
url = {https://bmvc2022.mpi-inf.mpg.de/BMVC2023/0462.pdf}
}
Copyright © 2023
The British Machine Vision Conference is organised by